Lihua Lei @lihua_lei_stat Twitter profile

Pinned Tweet

Lihua Lei

2 years

🚨Job talk thread🚨 Title: What Can *Conformal Inference* Offer to Statistics? Slides: Main points: (1) Conformal Inference can be made applicable in many #stats problems (2) There are lots of misconceptions about Conformal Inference (3) Try it! 1/n

4

54

296

Last Seen Profiles

@Carl_Bourne

@phoebe_kongwy

@0802onmark

@BinorRaja

@EduVG89

@HomoSap96578730

@stw_pdg

@daobase_ai

@Rebecca_Bellado

@Brightchigozie_

@jandakembangstw

@p41932848

@stratnewsint

@KarpasStev14445

@162310J

@JabulaniSi40934

@sclhenry

@JilatanJilbab

@cinemacollecti1

@jandakembangstw

@cmsadvocacy

@stw46

@nikocoliterus

@ranjan_ronnie

@nothingbutnets

@NikilisRBX

@NaCl_HQ

@NorrisJuni36951

@SLSEngland

@JilatanJilbab

@jajatex

@hijabjilbab1

@Naveed_ISB

@OnTheMoveOTM

@MrKokoMaun86756

@enmto

Lihua Lei

@lihua_lei_stat

2 years

Thrilled to announce that I will be joining @StanfordGSB as a tenure-track Assistant Professor of Economics in July 2022. I’m so grateful to my advisors, family, and friends for the tremendous support!! Time to start a brand-new chapter with my amazing future colleagues!!

80

18

1K

Lihua Lei

@lihua_lei_stat

2 years

A thread on BUE & BLUE🔥 Gauss-Markov condition: 1) y=Xβ+ε 2) E[ε|X]=0 3) Cov(ε|X)=σ^2Σ Standard GM: 4) Σ=I The GM thm shows that OLS/GLS is BL(inear)UE. Hansen (’20) shows it holds for all unbiased est (inc. nonlinear) w/ an elegant proof (tilted density + Cramer-Rao) 1/n

Giuseppe Cavaliere

@CavaliereGiu

2 years

The OLS estimator is BUE, on top of being BLUE, under classic Gauss-Markov conditions — check out @BruceEHansen ’s new amazing #econometrica paper. Textbooks require some updating @jmwooldridge ? Question: how much a BUE can depart from linearity?

19

263

1K

4

93

456

Lihua Lei

@lihua_lei_stat

2 years

Poisson regression should have been more popular for nonnegative outcomes: (1) robust to misspecification as long as logE[Y|X]=Xb; see @jmwooldridge ’s book (2) coef can be interpreted as semielasticity (3) never produces negative predicts (4) easy to compute or L1/L2-penalize

Arpit Gupta

@arpitrage

2 years

Wow, really cool discussion that's both informative and entertaining. Why to use Poisson, rather than log(1+y) in your regression:

1

22

110

7

44

325

Lihua Lei

@lihua_lei_stat

2 years

Done with my last interview today! 🎉🎉 In the past 3 months, I’ve had one-on-one meetings with >200 professors and gone through >80 papers during the preparation. It’s such an amazing experience to learn tons of new things and sell my other works not covered in the job talk!

7

1

313

Lihua Lei

@lihua_lei_stat

4 months

Thrilled to share that I received the #NSF CAREER Award! 🎉 Grateful for the support and excited to continue my study of distribution-free inference in stats & econometrics. Huge thanks to @StanfordGSB and my friends for assistance in proposal writing!

35

1

251

Lihua Lei

@lihua_lei_stat

2 years

@pqblair This excellent review by @jondr44 , @pedrohcgs , @ambilinski , and @DavidPoe223 gives a list of R and stata packages.

2

51

215

Lihua Lei

@lihua_lei_stat

3 years

Super excited that our paper "conformal inference of counterfactuals and individual treatment effects" (w/ Emmanuel Candès) has been accepted by JRSSB!! I'm so grateful for the valuable feedback from our reviewers, seminar audience, and people who reach out to us.

5

3

161

Lihua Lei

@lihua_lei_stat

3 years

🚨Job market thread🚨 I’ll be on the #Stats job market this year. Check out my CV here In this thread, I will tweet my works in no particular order. Any comments and suggestions will be appreciated!

0

20

143

Lihua Lei

@lihua_lei_stat

3 years

#ShamelessSelfPromo #statstwitter Dear seminar/conference/reading-group organizers, if you need a speaker, please DM/email me. I’d love to share my works on conformal inference, causal inference, multiple testing, network clustering, optimization, and panel data econometrics.

0

23

116

Lihua Lei

@lihua_lei_stat

3 years

Check out our new work ! We propose a framework, inspired by conformal inference, that is able to control risk for any #MachineLearning algorithms in finite samples for iid data w/o distributional assumptions! #statstwitter #computervision #deeplearning #ai

Distribution-Free, Risk-Controlling Prediction Sets

While improving prediction accuracy has been the focus of machine learning in recent years, this alone does not suffice for reliable decision-making. Deploying learning systems in consequential...

arxiv.org

Anastasios Nikolas Angelopoulos

@ml_angelopoulos

3 years

Just arXiv'd "Distribution-Free, Risk-Controlling Prediction Sets" (RCPS)! RCPS gives formal finite-sample 'confidence intervals' for any predictor on any prediction task, e.g. AlphaFold protein structure predictions as below. 1/n

2

30

101

2

28

108

Lihua Lei

@lihua_lei_stat

5 months

A must-read for folks who worry about negative weights! In our paper (), we showed negative weights are no concern in design specifications for Two-Way-Fixed-Effects estimators either! The result holds for almost any design, not just for staggered rollouts

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

Peter Hull

@instrumenthull

5 months

Happy New Year! Kirill @Borusyak and I have a New (short) Paper on the infamous "negative weights" issue recently raised for TWFE and other popular OLS/IV specifications Here's an (even shorter) summary thread

3

85

396

0

18

107

Lihua Lei

@lihua_lei_stat

3 years

Deeply honored to have been selected as a Rising Star in Data Science @DSI_UChicago ! It’s wonderful to attend the workshop and talk with my amazing mentor @WillettBecca , friends in stats, and other participants.

Data Science Institute

@DSI_UChicago

3 years

Meet our Autumn 2021 cohort of Rising Stars in Data Science:

0

2

16

4

2

100

Lihua Lei

@lihua_lei_stat

3 years

IV can be used to identify: (1) ATE=ATT, assuming constant treatment effect (2) ATT under one-sided compliance (3) LATE under monotonicity (4) (partially) ATE or ATT when IV and treatment are discrete Any other setting where IV can identify certain causal effect? #EconTwitter

7

10

96

Lihua Lei

@lihua_lei_stat

2 years

@causalinf @jondr44 @instrumenthull The phrases “model-based” and “design-based” come from survey sampling, in which “model” means “assumptions on the population (outcomes)” and “design” means “how a unit is sampled”, instead of the “research design” in Econ (e.g., IV, DiD, RDD). 1/n

2

17

92

Lihua Lei

@lihua_lei_stat

3 years

🚨New working paper🚨w/ @ArkhangelskyD , Guido Imbens, and @xiaoman_luo . We propose the RIPW estimator that makes TWFE regression robust to treatment effect heterogeneity under general designs (incl. staggered adoption) I’ll walk through some details. 1/n

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

Dmitry Arkhangelsky

@ArkhangelskyD

3 years

1/5 (Reweighted) two-way regression strikes back in our new paper with Guido Imbens, @lihua_lei_stat , and Xiaoman Luo ()! Can (and should) be used whenever data allows us to meaningfully discuss the model for the treatment assignment.

1

26

143

3

29

81

Lihua Lei

@lihua_lei_stat

2 years

Online Causal Inference Seminar 2022 will kick off today at 8:30PT/11:30ET with an interview with Nobel Laureate Guido Imbens!！

OCIS

Online Causal Inference Seminar

sites.google.com

0

10

60

Lihua Lei

@lihua_lei_stat

3 years

Join me (today at 3:30pm pst) if you are curious about *what conformal inference can offer to statistics*

UW Statistics

@UWStat

3 years

Join our seminar today (10/15) at 3:30pm (PST) via Zoom as Dr. Lihua Lei gives his talk on "Conformal Inference of Counterfactuals and Time-to-event Outcomes". For more details, please see:

0

2

15

4

5

58

Lihua Lei

@lihua_lei_stat

3 years

This is so awesome!!! It's my greatest honor to be able to work with Guido over the past few years!!

The Nobel Prize

@NobelPrize

3 years

BREAKING NEWS: The 2021 Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel has been awarded with one half to David Card and the other half jointly to Joshua D. Angrist and Guido W. Imbens. #NobelPrize

194

5K

12K

0

56

Lihua Lei

@lihua_lei_stat

3 years

Check out our new work on conformalized survival analysis w/ @RenZhimei and Emmanuel Candès: Our method can wrap around any survival predictive algorithms and produce calibrated covariate-dependent lower predictive bounds (LPBs) on survival times. 1/n

1

14

56

Lihua Lei

@lihua_lei_stat

2 years

Interesting to see how this repo compares w/ the thought-provoking paper (by @Susan_Athey , Imbens, Metzger & Munro) that applies Generative Adverserial Networks (GAN) to generate synthetic data that emulates real data: Code:

GitHub - gsbDBI/ds-wgan: Design of Simulations using WGAN

Design of Simulations using WGAN. Contribute to gsbDBI/ds-wgan development by creating an account on GitHub.

github.com

Philip Vollet

@philipvollet

2 years

The Synthetic Data Vault (SDV) is a synthetic data generation ecosystem to easily learn single-table, multi-table, and time-series datasets to generate new synthetic data that has the same format and statistical properties as the original dataset.

5

64

301

0

7

52

Lihua Lei

@lihua_lei_stat

2 years

I just created the community “No Boundaries Statistics”! The goal is to share information and insights from across the statistical spectrum (applications, computing, data, methodology, theory, software, …). Anyone who is interested is welcome to join!

0

8

50

Lihua Lei

@lihua_lei_stat

2 years

@jmwooldridge @CavaliereGiu @BruceEHansen There is a class of nonlinear unbiased estimators called Linear Plus Quadratic estimators developed in the 90s.

3

4

50

Lihua Lei

@lihua_lei_stat

3 years

@gabrielpeyre Here are two bounds that are tight when A and B commute. They involve the proximity of eigenspaces of A and B. I believe they can be further improved.

1

0

49

Lihua Lei

@lihua_lei_stat

6 years

Excited to release our R package "adaptMT" on Adaptive P-value Thresholding (AdaPT), an extremely flexible framework to incorporate side information into multiple hypthesis testing with false discovery rate (FDR) control! @wfithian Check out our vignette

0

19

49

Lihua Lei

@lihua_lei_stat

2 years

Great to see discussions putting strict exogeneity (mostly in econometrics) and sequential ignorability (mostly in biostats) on the same table! Time to have MORE in-depth conversations b/t #EconTwitter , #Polisci , and #BioStats , and think about interesting middle grounds

Political Methodology

@PolMethSociety

2 years

The fantastic @chadhazlett is teaching us all about TSCS data!

0

7

53

0

3

48

Lihua Lei

@lihua_lei_stat

4 months

Mike Jordan’s interview on AI and economics in the beautiful Trieste by @barbara_vyrill Love what Mike said: “I spend half of a typical day trying to minimize the entropy and the other half trying to maximize the entropy”

Talking about AI with the Italian Michael Jordan

Prof. Michael Jordan offers his provocative thoughts on the blending of AI and economics and takes us on a tour of Trieste, a beautiful and grand city in nor...

www.youtube.com

1

6

43

Lihua Lei

@lihua_lei_stat

2 years

Conformal inference is often framed as a tool for uncertainty quantification. But it can also be viewed as controlling the “miscoverage risk” for a particular type of predictions (pred interval/set). We finally extend it to any bounded monotone risk and any type of predictions!

Anastasios Nikolas Angelopoulos

@ml_angelopoulos

2 years

I’m thrilled to announce Conformal Risk Control: a way to bound quantities other than coverage with conformal prediction. Check out the worked examples in CV and NLP! The best part is: it’s exactly the same algorithm as split conformal prediction🤯🧵1/5

2

29

148

1

7

43

Lihua Lei

@lihua_lei_stat

7 months

Check out the presentations from my group at @CODEConference ! We are eager to get your feedback! Friday -Poster: Inference for synthetic control via robust placebo tests (by Timothy Sudijono) -Sec E (2:40): Multiple A/B testing with always-valid e-values (by Will Hartog) 1/2

3

6

42

Lihua Lei

@lihua_lei_stat

3 years

Excited to talk about my recent work w/ Emmanuel Candès at @stitchfix_algo on reliable uncertainty quantification of counterfactuals and individual treatment effects via conformal inference! Paper link:

Conformal Inference of Counterfactuals and Individual Treatment Effects

Evaluating treatment effect heterogeneity widely informs treatment decision making. At the moment, much emphasis is placed on the estimation of the conditional average treatment effect via...

arxiv.org

Stitch Fix Algorithms

@stitchfix_algo

3 years

Come join us for the next installation of Algo-Hour with Dr. Lihua Lei. In this talk, Dr. Lihua Lei ( @lihua_lei_stat ) will tell us about heterogeneous treatment effects and interval estimates for counterfactuals

4

5

29

0

3

42

Lihua Lei

@lihua_lei_stat

3 years

(Conditional) parallel trend = (mean-)ignorability on level growth of baseline potential outcomes A useful shortcut to memorize modern DiD estimators (OR by Heckman et al '97, IPW by Abadie '05, DR by @pedrohcgs &Zhao '20), given the familiarity to cross-sectional ATT estimators

2

41

Lihua Lei

@lihua_lei_stat

2 years

A great use case of our Reshaped IPW-TWFE estimator 1) no parallel trend & no pretreatment data to justify it 2) randomized treatment paths (design-based inference w/ small samples) 3) heterogeneous unit/time effects 4) limited spatial-temporal carryover

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

scott cunningham

@causalinf

2 years

Yeah I know I said I was getting off but I have a problem and need an AA sponsor (Adele Anonymous). Here’s a post abt Netflix using did. Can’t remember if I posted it or not.

0

5

22

1

9

41

Lihua Lei

@lihua_lei_stat

2 years

Special thanks to my wife @xiaoman_luo , and my letter writers Peter Bickel, Emmanuel Candès, Will Fithian @wfithian , Guido Imbens, and Michael Jordan!

2

0

41

Lihua Lei

@lihua_lei_stat

2 years

Check out #ASSA2022 session "What Can AI Do in Economics" at 10AM EST on Sunday! I'm so thrilled to discuss the excellent paper by @VC31415 , Whitney Newey, & Vira Semenova. I'll talk about interesting connections b/t their work and Reinforcement Learning

0

4

40

Lihua Lei

@lihua_lei_stat

3 years

New paper alert🚨 #statstwitter Conformal inference, often framed as a technique to generate prediction intervals, is also a tool for out-of-distribution detection. We studied marginal/conditional conformal p-values for multiple testing with marginal/conditional error control 1/n

Stephen Bates

@stats_stephen

3 years

Conformal inference gives rigorous outlier/out-of-distribution detection. We show how to control FDR with conformal p-values -- even though they are dependent, they satisfy the PRDS property! With E. Candès, @lihua_lei_stat , Y. Romano, and M. Sesia

0

10

39

1

7

40

Lihua Lei

@lihua_lei_stat

2 years

Honored to present my work w/ @wfithian on AdaPT in this year’s @RoyalStatSoc Journal Webinar! First time presenting while losing voice😰Proud of myself💪 Video: Paper: (free access) R package:

adaptMT: Adaptive P-Value Thresholding for Multiple Hypothesis Testing with Side Information

Implementation of adaptive p-value thresholding (AdaPT), including both a framework that allows the user to specify any algorithm to learn local false discovery rate and a pool of convenient functi...

cran.r-project.org

1

3

38

Lihua Lei

@lihua_lei_stat

3 years

Excellent point made by @edwardhkennedy and @jondr44 . The amount of information that can be reliably learned from data always relies on underlying assumptions. IMHO, it's much more informative to show the result-assumption tradeoff explicitly whenever possible.

Edward Kennedy

@edwardhkennedy

3 years

This is why I love combining bds *with* sensitivity analysis. You get the whole picture: pt identification @ one extreme, agnostic bounds at the other Best of both worlds & middle ground in between! Nice example in @bonv3 % unmeasured confounding paper:

2

1

41

1

3

36

Lihua Lei

@lihua_lei_stat

3 years

Curious about how to wrap around any black-box algorithm to get trustworthy counterfactual inference? Come to my talk today at 2PM (pacific time) @stitchfix_algo Zoom link 👇

0

6

37

Lihua Lei

@lihua_lei_stat

3 years

Great thread on Poisson regression! It's also great for conditional variance estimation: 1) Est. conditional mean m(x) 2) Pois. reg. e=(Y - m(X))^2 on X 3) Est. Var(Y|x). as exp(x'b) More stable than lm(log(e)~X), and better than lm(e~X) which can't guarantee positive estimates

Jeffrey Wooldridge

@jmwooldridge

3 years

Poisson regression can get one so far with so little trouble, why do so many still resist? Especially with panel data. It’s too bad we can’t give it another name to reflect the fact that its a fully robust estimator of conditional mean parameters.

51

272

2K

0

1

37

Lihua Lei

@lihua_lei_stat

3 years

Want to assess population overlap/positivity in observational studies w/ an inconsistent propensity score est.? Come to my talk w/ @AviFeller tomorrow 10:20-10:45 PT at #CausalAssumptions #ICML2021 w/ @alexdamour , Peng Ding, @AviFeller , and Jas Sekhon

0

7

36

Lihua Lei

@lihua_lei_stat

3 years

Check out my contributed talk "Calibrated Out-of-Distribution Detection with Conformal P-values" at 4PM EST #UDL2020 #ICML2021 I'll talk about using conformal inference to calibrate ANY score-based OOD detection method for single/multiple testing Paper:

Testing for Outliers with Conformal p-values

This paper studies the construction of p-values for nonparametric outlier detection, taking a multiple-testing perspective. The goal is to test whether new independent samples belong to the same...

arxiv.org

Balaji Lakshminarayanan

@balajiln

3 years

1/n Attending #ICML2021 ? Join us for our workshop on "Uncertainty and Robustness in Deep Learning" @icmlconf this Friday (July 23) 9am-5pm ET! ICML link: Co-organized w/ @DanHendrycks @SharonYixuanLi @latentjasper @tdietterich @csilviavr @sebnowozin

3

65

292

2

4

36

Lihua Lei

@lihua_lei_stat

3 years

New working paper We found a deep connection b/t PAC-learning &familywise error rate (FWER) control, two nearly non-overlapping areas. Tools used by drug companies (graphical approach for FWER) are very useful for computer vision! #mltwitter #statstwitter

Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control

We introduce a framework for calibrating machine learning models so that their predictions satisfy explicit, finite-sample statistical guarantees. Our calibration algorithms work with any...

arxiv.org

Anastasios Nikolas Angelopoulos

@ml_angelopoulos

3 years

Thrilled to share Learn then Test, a tool to calibrate any model to control risk (eg. IOU, recall in object detection). No assns on model/data. See arXiv + Colab ✍️w/ @stats_stephen , E.J. Candes, M.I. Jordan, @lihua_lei_stat ! 🧵1/n

4

24

89

1

7

32

Lihua Lei

@lihua_lei_stat

10 months

Congrats!!! Hugely well deserved!

COPSS

@COPSSNews

10 months

👏 The secret is out! The recipient of the 2023 COPSS President’s Award is… Dr. Ryan Tibshirani from @UCBerkeley !! #JSM2023

0

14

111

0

32

Lihua Lei

@lihua_lei_stat

2 years

Paper accepted by #SIMODS journal🍺! The #optimization lit has tons of rate-optimal algorithms. Time to seriously think about how to make them adaptive (e.g., achieving the rate for strongly convex functions w/o knowing it’s strongly convex)

Adaptivity of Stochastic Gradient Methods for Nonconvex Optimization

Adaptivity is an important yet under-studied property in modern optimization theory. The gap between the state-of-the-art theory and the current practice is striking in that algorithms with...

arxiv.org

Lihua Lei

@lihua_lei_stat

4 years

Our new work on adaptivity of stochastic gradient methods for smooth nonconvex optimization. . We propose Geom-SARAH that achieves unusually strong adaptivity to both target accuracy and PL constant without knowing them! Main technique: Geometrization! 1/5

1

14

4

2

31

Lihua Lei

@lihua_lei_stat

2 years

So proud that our International Selective Inference Seminar (ISSI) has hosted 75 seminars from 4/16/2020!! Many thanks to all amazing speakers, discussants, participants, and co-organizers Rina Barber, @wfithian , and Daniel Yekutieli. Join us next year!

0

3

31

Lihua Lei

@lihua_lei_stat

2 years

Shout out to @AsjadNaqvi who did an amazing job curating a list of R and stata packages for DiD methods! Also an important lesson for me that we should give enough credit to those fascinating unpublished work and code repo when sharing/tweeting!

Asjad Naqvi

@AsjadNaqvi

2 years

And that is how unpublished work disappears from attribution in the second and third waves of online sharing. It took almost a year to get these packages and descriptions in order.

6

16

170

1

4

31

Lihua Lei

@lihua_lei_stat

2 years

Check out my talk if you’re curious about how to make the TWFE estimator robust to (1) arbitrary violation of parallel trends (2) unrestricted heterogeneity in unit/time effects by modeling the assignment mechanism and reweighting the TWFE regression

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

Peter Hull

@instrumenthull

2 years

@borusyak @paulgp @stephencoussens This workshop follows in the footsteps of the massively successful DD workshop, by the great (& similarly very online) @pedrohcgs , @CdeChaisemartin , @jondr44 , Brant Callaway, and @lihua_lei_stat You can watch a recording of this earlier workshop here

0

10

40

2

7

31

Lihua Lei

@lihua_lei_stat

2 years

Join us tomorrow if you are interested in conformal inference + causal inference + semiparametrics! #ISSI

Lihua Lei

@lihua_lei_stat

2 years

#ISSI This Thursday (8:30AM PT) we will have Yachong Yang talking about “Double robust prediction with covariate shift” (joint w/ Prof. Arun Kumar Kuchibhotla & Prof. Eric Tchetgen Tchetgen), followed by a discussion by Prof. James Robins.

0

5

16

0

9

31

Lihua Lei

@lihua_lei_stat

1 month

Need help from experts on #analysis #PDE ! (1) How does the constant in Gagliardo-Nirenberg inequality (C in the screenshot) depend on dimension n? (2) How to extend this particular one to fractional Sobolev spaces? Brezis-Mironescu doesn’t look like so (or I could be wrong)

4

30

Lihua Lei

@lihua_lei_stat

2 years

I’m so proud that my job talk convinced many folks to read more about Conformal Inference. There are tons of exciting methodological questions (distribution shifts, dependence, conditional coverage, …) and real-world applications. It’s a very🔥area now! Come and join us! 6/6

3

0

30

Lihua Lei

@lihua_lei_stat

3 years

Check out my talk if you are curious about experimental designs for TWFE estimators! I’ll be talking about a very surprising result. joint w/ @ArkhangelskyD , Guido Imbens, and @xiaoman_luo Paper link:

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

Causalscience.org

@causal_science

3 years

We’re looking forward to welcome @IavorBojinov , @yixinwang_ , @lihua_lei_stat , Zhenyu Zhao and @totteh in Session 4! Join us at 4:40pm (CET). #CDSM21 #CausalInference

0

2

9

0

6

30

Lihua Lei

@lihua_lei_stat

2 years

So much fun reading this excellent paper together w/ the one by @paulgp -Sorkin-Swift (AER’20) on shift-share IV! Great example that model≠estimator. Same 2SLS w/ the same Bartik IV can be justified using diff. sources of randomness (shifts for the former, shares for the latter)

Xavier Jaravel

@XJaravel

2 years

Great to see our paper on quasi-experimental shift-share designs with @borusyak @instrumenthull coming out in print this month @RevEconStudies ! 👇👇👇👇👇👇👇👇

0

31

280

0

2

28

Lihua Lei

@lihua_lei_stat

2 years

Great point on GLMs! Li, Morgan, and Zaslavsky (’18) showed that logistic regression yields a form of exact covariate balancing even if the logistic model is completely wrong. A similar result was derived by Imai and Ratkovic (’12)

Balancing Covariates via Propensity Score Weighting

Covariate balance is crucial for unconfounded descriptive or causal comparisons. However, lack of balance is common in observational studies. This article considers weighting strategies for balanci...

www.tandfonline.com

Dmitry Arkhangelsky

@ArkhangelskyD

2 years

@lihua_lei_stat @jmwooldridge Poisson regression is a great thing! But GLMs in general are great tools, no? For example, dual for balancing problems (e.g., synth) can typically be interpreted as penalized GLM

1

0

8

0

6

27

Lihua Lei

@lihua_lei_stat

2 years

Our International Seminar of Selective Inference (ISSI) has a very broad theme On Thursday (8:30AM PT), we will have Richard Berk talking about algorithmic fairness, optimal transport, and conformal inference, followed by a discussion by Emmanuel Candès

0

4

27

Lihua Lei

@lihua_lei_stat

3 years

An important note by @omaclaren that "identifiability" doesn't always imply "estimability with finite uncertainty". Bahadur-Savage ('56) show it's impossible to get valid nontrivial confidence interval even for MEAN when the info about the dist. is limited

The Nonexistence of Certain Statistical Procedures in Nonparametric Problems

projecteuclid.org

Oliver Maclaren

@omaclaren

3 years

@stephensenn @AdanZBecerra1 @ProfMattFox In causal inference identification is about the infinite data setting, so yup they just consider point estimates (functionals of full population distribution). It’s easy enough to show that identifiability does not imply estimability with finite uncertainty [insert usual link]

1

2

8

1

3

27

Lihua Lei

@lihua_lei_stat

3 years

#statstwitter #CausalInference #EconTwitter What are the minimal assumptions on the estimates of nuisance parameters under which valid inference of ATE based on the Augmented-IPW estimator is possible?

2

5

26

Lihua Lei

@lihua_lei_stat

2 years

@jmwooldridge @CavaliereGiu @BruceEHansen Just found that @BruceEHansen cited Koopmann (’82) and Gnot et al. (’92), which showed that all unbiased and *equivariant* estimators for linear models (Y->Y+Xγ => \hat{β}->\hat{β}+γ) are certain Linear-Plus-Quadratic estimators.

1

4

23

Lihua Lei

@lihua_lei_stat

4 years

Excited to present my work with Emmanuel Candès on conformal inference of counterfactuals and individual treatment effects at @CODEConference tomorrow (Thursday, 8:15AM PT, parallel session 1A)! Paper link: Also look forward to other interesting talks!

Conformal Inference of Counterfactuals and Individual Treatment Effects

Evaluating treatment effect heterogeneity widely informs treatment decision making. At the moment, much emphasis is placed on the estimation of the conditional average treatment effect via...

arxiv.org

MIT IDE

@mit_ide

4 years

Have you registered for the 2020 @CODEConference ? Join leading researchers from various scientific disciplines including economics, computer science & sociology. Organized by @MIT leaders: @sinanaral , @deaneckles , @alex_pentland @johnjhorton Register:

0

2

11

1

25

Lihua Lei

@lihua_lei_stat

1 year

Thrilled to present in the NISS-Merck Meet-up! Check out my talk at 10:30PT/1:30ET if you are interested in how to generalize conformal prediction to conformal risk control.

NISS

@NISS_DataSci

1 year

Join us tomorrow! Speakers: Dr. Alexander Gammerman from @RoyalHolloway , Dr. Yao Xie from @GeorgiaTechISyE , Dr. Matteo Sesia from @USCMarshall , and Dr. @Lihua Lei @lihua_lei_stat from @StanfordGSB Moderator: Dr. Yuting Xu from @Merck RESVP:

0

4

8

1

4

24

Lihua Lei

@lihua_lei_stat

2 years

Good writing is so important for methodologists/theorists. It’s especially challenging for those like me whose native tongue is non-Germanic. I’m so lucky to learn from extraordinary writers like @wfithian , Michael Jordan, and Emmanuel Candès, whose writing skills are remarkable.

Winston Lin

@linstonwin

3 years

"One should take advantage of the English language, and not just rely purely on mathematical symbols."

1

6

21

1

0

24

Lihua Lei

@lihua_lei_stat

2 years

Got my booster today! Feel so lucky to have these two holiday weeks to take a breath during the job market season. Perfect time for JMCs to get boosters without fear of side effects!

0

24

Lihua Lei

@lihua_lei_stat

4 years

Excited to release our Shiny app on survival analyses of publicly available datasets on #Covid_19 @RenZhimei @xiaoman_luo ! . We fit Kaplan-Meier curves on four populations stratified by age or sex to estimate the survival prob. for a certain number of days.

1

9

23

Lihua Lei

@lihua_lei_stat

1 year

@jondr44 This is a special case of the survivor average causal effect E[Y(1) - Y(0) | S(1) >c, S(0)>c) with Y(t)=S(t), c=0. It was introduced in Zhang and Rubin (2003) which also derived the bounds. Part of their motivation is same as yours.

1

23

Lihua Lei

@lihua_lei_stat

2 years

@causalinf @jondr44 @instrumenthull It’s perhaps more precise to frame “model-based” as “outcome-based”, and “design-based” as “treatment-based”. The former relies on conditions like linearity, SEM, parallel trend, continuity of PO; the latter relies on conditions like random timing, propensity score models. 5/n

1

3

23

Lihua Lei

@lihua_lei_stat

2 years

One popular question in my interviews is “why is conformal inference called conformal?” Long answer short, it has nothing to do w/ “conformal mappings” in geometry, but roughly means “conformity to data”. Vladimir Vovk, one of the inventors, clarified it👇

0

1

23

Lihua Lei

@lihua_lei_stat

2 years

Excited to present (~10AM PST) some really surprising properties of the two-way fixed effect (TWFE) estimator with fully heterogeneous treatment effects and no parallel trend. joint w/ @ArkhangelskyD , Guido Imbens, and @xiaoman_luo Paper link:

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

Pedro H. C. Sant'Anna

@pedrohcgs

2 years

Hey #EconTwitter and otherwise! On December 15th, we will have a webinar about recent advances in DiD methods. Speakers are @lihua_lei_stat , @CdeChaisemartin , @jondr44 and #twitterless Brantly Callaway. More info: Will be fun! See y'all there.

4

61

227

0

6

23

Lihua Lei

@lihua_lei_stat

2 years

Partial identification meets shape constraints. Very cool paper!!

Shosh Vasserman

@shoshievass

2 years

Very excited to share this new paper with the fantastic @ZiYangKang , out on NBER today. Thread 👇 with an overview.

2

44

311

0

21

Lihua Lei

@lihua_lei_stat

1 month

Need help (again!) from experts on #stats learning theory: How does the constant in the L2 minimax rate O(n^-s/(2s+d)) for estimating Holder functions (conditional mean or density) depend on dimension d and Holder order s? What is hidden in the big-O?

4

6

22

Lihua Lei

@lihua_lei_stat

3 years

Is there any real-world example where invariant learning (e.g., IRMv1) achieves good invariance (i.e., outcome independent of environment given the prediction) with a moderate number of environments (say, 10 or 20)? #mltwitter #statstwitter

1

6

22

Lihua Lei

@lihua_lei_stat

4 years

Check out my talk on top-down hierarchical spectral clustering of networks A line of works w/ amazing collaborators. Join my live discussion (8/25 12-2 PT, room 4 ) if you’re interested #worldsymposium2020 @BernoulliSoc @InstMathStat

3

20

Lihua Lei

@lihua_lei_stat

3 years

@96Henrique @elben @analisapackham You might be interested in our paper (w/ @ArkhangelskyD , Guido Imbens, and @xiaoman_luo ) for a "natural evolution" of TWFE. We proposed a weighted TWFE estimator that resolves many issues of TWFE (even w/o parallel trend in some cases).

Design-Robust Two-Way-Fixed-Effects Regression For Panel Data

We propose a new estimator for average causal effects of a binary treatment with panel data in settings with general treatment patterns. Our approach augments the popular two-way-fixed-effects...

arxiv.org

0

5

19

Lihua Lei

@lihua_lei_stat

3 years

This counterfactual interpretation of vaccine efficacy is so neat!!

Jaffer Zaidi

@jzaidi14

3 years

@ProfMattFox @kaz_yos (1) Among the individuals randomized to the placebo arm that ended up getting covid (say 100), at the very least 95 of these 100 covid patients would have been covid-free if given the vaccine. (In potential outcomes, P(Y(1)=0|Y(0)=1) >= VE =0.95). It's a lower bound.

1

3

19

2

20

Lihua Lei

@lihua_lei_stat

4 years

A fascinating panel at #CODECON20 ! As a statistician in academia, it is tremendously valuable for me to learn from a META-PLATFORM like this! @deaneckles Have you considered organizing a weekly online webinar on digital experimentation as an extension of CODE?

MIT IDE

@mit_ide

4 years

Platform giants, @Netflix , @Airbnb , and @Facebook explain how they use complex digital experimentation in their ad tracking, app design testing, and analytics at the #CODECON2020 Practitioners Panel with @ronnyk @lilidworkin and Jeffrey Wong

2

4

34

3

1

19

Lihua Lei

@lihua_lei_stat

3 years

I’ll be talking about conformal inference for time-to-event data at the IFDS-MADLab Workshop tomorrow 8:15-9AM PST We adapted the weighted conformal inference to generate calibrated and doubly robust lower predictive bounds for actual survival times.

Lihua Lei

@lihua_lei_stat

3 years

Check out our new work on conformalized survival analysis w/ @RenZhimei and Emmanuel Candès: Our method can wrap around any survival predictive algorithms and produce calibrated covariate-dependent lower predictive bounds (LPBs) on survival times. 1/n

1

14

56

0

1

19

Lihua Lei

@lihua_lei_stat

3 years

Excellent thread on covariate-adaptive false discovery rate control! Wolfgang and Nikos opened a new line of research in multiple testing, including mine w/ @wfithian and others: AdaPT (for generic setting), STAR (for logical constraint), and BONuS (for multivar. test statistics)

Wolfgang Huber 🇺🇦

@wolfgangkhuber

3 years

1/ In this tweetorial, I present how multiple testing can be made more powerful by using freely available informative side information (“covariate-powered multiple testing”), as described in a new paper @NikosIgnatiadis ().

2

45

115

0

2

19

Lihua Lei

@lihua_lei_stat

3 years

@jiafengchen42 I think you can't construct uncountably many iid U([0,1])s if the probability space is "nice" (e.g. Polish). Otherwise, you would obtain an uncountable orthonormal basis for a separable Hilbert space (the set of bounded rvs equipped with the inner product as the covariance).

1

18

Lihua Lei

@lihua_lei_stat

3 years

This is an amazing workshop!

Anastasios Nikolas Angelopoulos

@ml_angelopoulos

3 years

Announcing the first-ever Workshop on Distribution-Free Uncertainty Quantification (DF UQ) at @ICML2021 . About UQ without any assumptions on the model or #data distribution. All r welcome to submit talks/papers! It's gonna be AWESOME!🧵1/6 #AI #ML

3

44

110

1

17

Lihua Lei

@lihua_lei_stat

2 years

Can’t agree more! From my own exp. as a methodologist, publishing a paper is often the start of a project. Explanation, translation, and implementation that follow are at least equally important. Don’t blame others for not using new methods. Try convincing them to do so instead!

Paul Goldsmith-Pinkham

@paulgp

2 years

Francesca makes an amazing comment regarding this that I wish got pushed further. Imagine econometrician ref thinks methods have already been done. Applied ref thinks empirics has been done but method is cool. My point: why doesn’t applied ref know about the method then???

1

15

0

18

Lihua Lei

@lihua_lei_stat

3 years

Surprising fact about panel experiment designs #EconTwitter #statstwitter For an RCT w/ staggered rollouts, what randomization scheme (prob. of each adopting time) guarantees TWFE estimator to be consistent for ATE (over units and time)? Uniform dist.? NO if >2 periods 1/3

2

7

18

Lihua Lei

@lihua_lei_stat

2 years

#ISSI This Thursday (8:30AM PT/11:30AM ET), we will have Yonghoon Lee talking about “Distribution-free inference for regression: discrete, continuous, and in between” (), followed by a discussion by @YingJin531

0

4

17

Lihua Lei

@lihua_lei_stat

1 year

@CasualBrady Check out our paper (w/ @Alexdamour , @AviFeller , Peng Ding, and Jas Sekhon)! Turns out positivity violation can be tested even if the propensity score is completely wrong. We only assume iid for validity, and it has high power when the ps model is good.

1

0

17

Lihua Lei

@lihua_lei_stat

3 years

Amazing result for binary IV + binary treatment!!! The IV (2SLS) estimand can still be interpreted as a causal effect (LATT) even if there is an unmeasured confounder between the “IV” and the treatment, as long as neither defiers nor always-takers exist.

Kevin Chen

@jiafengkevinc

3 years

Calculated the IV estimand in the binary case:

6

5

28

0

2

17

Lihua Lei

@lihua_lei_stat

2 years

@AsjadNaqvi Sorry for that, Asjad! I should have read the review paper more carefully. Your website is absolutely amazing! I’ll make sure to credit it in my future work (and tweets).

1

0

17

Lihua Lei

@lihua_lei_stat

3 years

2020’s “year in reviews”: I did 20 journal reviews (excluding reviews of revisions), and 28 conference reviews. I’ve learned a lot from those papers. Many thanks to the authors for their wonderful works and the AEs for inviting me to review!!

Gautam Kamath

@thegautamkamath

3 years

2020's "year in reviews": I personally did 30 conference reviews, 14 workshop reviews, and 1 journal review. Down from last year (38/11/1), but more papers overseen overall, being a core PC member/area chair for ICALP, ICML, RANDOM, ALT, and ICLR. Anyone else care to share #'s?

2

0

47

0

1

16

Lihua Lei

@lihua_lei_stat

2 years

Very useful tips for discussants 👏 For junior researchers who want to see examples on 5-10min discussions, check out the recordings from our International Seminar of Selective Inference (ISSI) We have had 75 great discussants in the past!

David Evans

@DaveEvansPhD

2 years

Ten tips for discussants at academic conferences (I was preparing discussant comments today and went back to this guidance from @CBlatts .)

9

128

595

0

2

16

Lihua Lei

@lihua_lei_stat

2 years

#ISSI This Thursday (8:30AM PT) we will have Yachong Yang talking about “Double robust prediction with covariate shift” (joint w/ Prof. Arun Kumar Kuchibhotla & Prof. Eric Tchetgen Tchetgen), followed by a discussion by Prof. James Robins.

0

5

16

Lihua Lei

@lihua_lei_stat

4 years

Check out my talk at #CDSM20 on conformal inference of counterfactuals and individual treatment effects at 8am PT!

Causalscience.org

@causal_science

4 years

Looking forward to a great third session of the day #CDSM20 Galen from @UW , @lihua_lei_stat from @Stanford , @_ignacio82 from @Google and Chad Hazlett from @UCLA #Causality #MachineLearning #DataScience #Econometrics #EconTwitter #MgmtTwitter 3BookofWhy

0

2

4

1

15

Lihua Lei

@lihua_lei_stat

2 years

Check out this awesome repo by @predict_addict if you are interested in conformal inference!

Valeriy M., PhD, MBA, CQF

@predict_addict

2 years

@lihua_lei_stat @lihua_lei_stat thank you for sharing slides, now featured on Awesome Conformal Prediction

1

3

6

0

3

15

Lihua Lei

@lihua_lei_stat

3 years

Check out Emmanuel Candès' talk today at 8:05PM PST on recent developments on conformal inference! #DFUQ2021 #ICML2021

Anastasios Nikolas Angelopoulos

@ml_angelopoulos

3 years

🔥 The second session of ICML DFUQ'21 is happening at 4:10pm PST 🔥 This half features invited talks from Jing Lei, Kilian Weinberger, and Emmanuel Candes. Then we will have two wonderful spotlight sessions featuring... 1/2

1

3

21

1

2

15

Lihua Lei

@lihua_lei_stat

2 years

Conformal Inference was designed for generating prediction intervals with guaranteed coverage in standard #ML problems. Nevertheless, it can be modified to be applicable in ✔️Causal inference ✔️Survival analysis ✔️Election night model ✔️Outlier detection ✔️Risk calibration 2/n

2

3

15

Lihua Lei

@lihua_lei_stat

3 years

#ASSA2021 #EconTwitter #statstwitter The “Econometrics Session in Honor of Gary Chamberlain” is AMAZING!! The discussion of @guido_imbens ’s talk is so enlightening @metrics52 @Susan_Athey @jmwooldridge . It is paramount to clarify *what is random* for claims involving uncertainty.

0

2

15

Lihua Lei

@lihua_lei_stat

4 years

Our new work on adaptivity of stochastic gradient methods for smooth nonconvex optimization. . We propose Geom-SARAH that achieves unusually strong adaptivity to both target accuracy and PL constant without knowing them! Main technique: Geometrization! 1/5

Adaptivity of Stochastic Gradient Methods for Nonconvex Optimization

Adaptivity is an important yet under-studied property in modern optimization theory. The gap between the state-of-the-art theory and the current practice is striking in that algorithms with...

arxiv.org

1

14

Lihua Lei

@lihua_lei_stat

2 years

#ISSI This Thursday (8:30AM PT) we will have Prof. Leying Guan talking about “Localized Conformal Prediction” (), followed by a discussion by Prof. Rafael Izbicki.

2

4

14

Lihua Lei

@lihua_lei_stat

2 years

Very neat proof on implicit bias of GD/SGD for generalized linear objectives! It generalizes the beautiful result (Prop 1 of ) that GD/SGD initialized at 0 for Least Squares converges to the minimum-norm solution (or the ridge estimator with penalty -> 0+)

Surprises in High-Dimensional Ridgeless Least Squares Interpolation

Interpolators -- estimators that achieve zero training error -- have attracted growing attention in machine learning, mainly because state-of-the art neural networks appear to be models of this...

arxiv.org

Fabian Pedregosa

@fpedregosa

2 years

January’s nugget is out! This is a short proof characterizing the implicit bias of gradient-based methods for regression losses: This is one of my favorite optimization proofs, as the technique looks like no other.

5

28

139

1

14

Lihua Lei

@lihua_lei_stat

4 years

Very cool application of conformal inference! Reliable uncertainty quantification is more appealing than throwing a model and hoping for the best

John Cherian

@jjcherian

4 years

On election night, The Washington Post will be using a model @lbronner and I built that relies on conformal inference to output intervals for likely vote totals in crucial swing states. I think that's pretty exciting!

3

6

49

0

3

14

Lihua Lei

@lihua_lei_stat

2 years

@causalinf @jondr44 @instrumenthull For example, are doubly robust estimators model-based or design-based? What about placebo tests for synthetic control? What about local randomization tests for RDD? To me, they are both model- and design-based, bc they involve assumptions on both outcomes and assignments. 4/n

1

0

14

Lihua Lei

@lihua_lei_stat

2 years

#ISSI This Thursday (8:30AM PT) we will have Neil Xu talking about “Post-selection inference for e-value based confidence intervals” (joint w/ Prof. Aaditya Ramdas & Prof. Ruodu Wang), followed by a discussion by Dr. Zhimei Ren @RenZhimei

0

3

14

Lihua Lei

@lihua_lei_stat

2 years

Really enjoy @rkoenker ’s fascinating talk on applying NPMLE-EB to Bradley-Terry model for ranking! #ASSA2022 They apply it to ranking 86 journals in #Econometrics & #stats . Glad to see my fav stats journals (JRSS-B, AoS, Biometrika, JASA) and prob journal (PTRF) all rank the top!

3

0

13

Lihua Lei

@lihua_lei_stat

4 years

Super interesting thread on GP and BayesOp! BayesOp is the first topic I worked on at Berkeley (w/ Prof. Cari Kaufman). Beyond the elegance and flexibility of GP, I was impressed by how useful uncertainty quantification is to balance exploration and exploitation (similar to UCB)

Yisong Yue

@yisongyue

4 years

I've found GPs to be very useful in real-world applications where data is scarce and uncertainty quantification is useful for (sequential) decision making. Thread below 🧵👇 1/

1

10

65

0

2

13

Lihua Lei

@lihua_lei_stat

3 years

Check out the talk by Samuel Horváth on our joint work w/ @peter_richtarik and Michael Jordan We proposed an algorithm for smooth nonconvex optimization that is provably adaptive to the target accuracy and Polyak-Łojasiewicz constant using "Geometrization"

Adaptivity of Stochastic Gradient Methods for Nonconvex Optimization

Adaptivity is an important yet under-studied property in modern optimization theory. The gap between the state-of-the-art theory and the current practice is striking in that algorithms with...

arxiv.org

Quanquan Gu

@QuanquanGu

3 years

1/4 Want to learn more about optimization for ML after @NeurIPSConf main conference? Welcome to our OPT2020 workshop on optimization for machine learning on Friday (in 6 hours!). Friday, December 11th, 2020 EST time 06:00 AM -- 19:00 PM Schedule:

1

24

124

0

2

13